Minority oversampling framework for class imbalance problem

Abstract

This dissertation enhances oversampling techniques which are used in a class imbalance problem. Several weaknesses of existing oversampling techniques are investigated and the minority oversampling framework is suggested to overcome these weaknesses and improves the classification performances. This dissertation provides the framework which contains three research works that deal with different aspects of existing oversampling techniques. The first work is Relocating Safe-level SMOTE (RSLS) to avoid conflicted synthetic instances near majority instances. The second work is Adaptive Neighbor SMOTE (ANS) which provides the dynamic number of nearest neighbors in SMOTE algorithm. The final work is the minority outcast handling process with 1-nearest neighbor to handle noises of positive instances in the dataset for improving the classification performance. This minority outcast handling process is augmented into RSLS and ANS to boost their accuracies. The experimental results on 14 benchmark datasets and 5 classifiers confirm that both oversampling techniques with minority outcast handling outperform other oversampling techniques in most datasets under three performance measures; F-measure, geometric mean and adjusted geometric mean. Wilcoxon sign ranked test is conducted to verify that the improvements caused by these two oversampling techniques are statistically significant.